Optimisation of Lane-Level Dynamic Traffic Control Strategy Based on Bidirectional Adaptive Gated Graph Convolutional Network and Deep Reinforcement Learning

Abstract

This paper innovatively proposes a lane-level dynamic traffic control strategy optimisation method integrating the bidirectional adaptive gated graph convolutional network (Bi-AGGCN) and deep reinforcement learning (DRL). The core innovation lies in three aspects. First, Bi-AGGCN is introduced to precisely capture the spatiotemporal dependencies of traffic flow by simultaneously considering forward and backward information, overcoming the limitations of traditional unidirectional models. Second, an improved deep Q-network (DQN) algorithm is adopted, incorporating a dual network structure, experience pool sampling strategy, and dominance function, which effectively enhances the learning speed and estimation accuracy of the value function. Third, the combination of Bi-AGGCN and DRL enables the framework to automatically adjust traffic signal parameters based on real-time traffic flow, achieving dynamic optimisation of traffic flow. Experimental results indicate that compared with traditional timed signal control (FTSC), fast Q-learning (FQ learning), and modified DQN (M-DQN) algorithms, the proposed Bi-AGGCN + DRL model demonstrates significant advantages. In the experiment, the traffic flow of this model reaches 2600 pcu/h, the delay time is reduced to 90 s, the lane-level control response speed is shortened to 5 s, and the average lane speed is increased to 110 km/h. This verifies the efficiency and feasibility of the model in lane-level traffic control, providing feasible technical support and optimisation directions for the traffic management of actual highways.