The development and application of autonomous vehicles bring great changes to urban traffic management and control. As one of the bottlenecks to improve transportation efficiency, intersection management plays an important role in the urban city. When the dynamic control method in different cases is determined, the key of autonomous intersection management problem is to search the passing orders for approaching connected and automated vehicles (CAVs). The paper proposed a framework based on convolutional neural network to predict different passing orders’ total time consumption. Thus, the best passing order with the lowest time consume can be chosen as the optimal solution. Then continuous-time optimal control can be carried out on CAVs. Meanwhile, sequential model-based algorithm configuration technique is used for neural network training. Simulation results exported from Simulation of Urban Mobility (SUMO) indicate that the proposed method outperforms actuated signal control and first come first serve strategy. The average delay of the proposed method can decrease by 42.40%–73.05% compared with actuated signal control and 2.95%–55.29% compared to first come first serve strategy. Moreover, it can increase average speed by more than 20% compared with the other two methods. The proposed method can significantly reduce the computation time comparing with the original planning-based strategy. At last, the framework can be applied to other regression tasks like vehicle emissions, then different optimization targets can be estimated to get better solutions faster.

自动驾驶汽车的发展和应用给城市交通管控带来了巨大的变化。作为提高交通效率的瓶颈之一，交叉口管理在城市中发挥着重要作用。当确定不同情况下的动态控制方法时，自动交叉口管理问题的关键是搜索接近的联网自动车辆（CAV）的通行顺序。论文提出了一种基于卷积神经网络的框架来预测不同传递订单的总时间消耗。因此，可以选择时间消耗最低的最佳通过顺序作为最优解。然后可以对 CAV 进行连续时间优化控制。同时，基于序列模型的算法配置技术用于神经网络训练。从城市交通仿真（SUMO）导出的仿真结果表明，所提出的方法优于驱动信号控制和先到先服务策略。与驱动信号控制相比，所提出方法的平均延迟可降低 42.40%–73.05%，与先到先服务策略相比，平均延迟可降低 2.95%–55.29%。而且，与其他两种方法相比，它可以将平均速度提高20%以上。与原始的基于规划的策略相比，所提出的方法可以显着减少计算时间。最后，该框架可以应用于其他回归任务，例如车辆排放，然后可以估计不同的优化目标以更快地获得更好的解决方案。