测绘学报 ›› 2020, Vol. 49 ›› Issue (12): 1630-1639.doi: 10.11947/j.AGCS.2020.20190516

• 地图学与地理信息 • 上一篇    下一篇

基于深度强化学习的电动出租车运营优化

叶浩宇1, 涂伟2,3,4,5, 叶贺辉6, 麦可2,3,4, 赵天鸿2,4, 李清泉1,2,3,4   

  1. 1. 武汉大学测绘遥感信息工程国家重点实验室, 湖北 武汉 430079;
    2. 深圳大学建筑与城市规划学院城市空间信息工程系, 广东 深圳 518060;
    3. 人工智能与数字经济广东省实验室(深圳), 广东 深圳 518060;
    4. 广东省城市空间信息工程重点实验室, 广东 深圳 518060;
    5. 自然资源部大湾区地理环境监测重点实验室, 广东 深圳 518060;
    6. 闽江学院软件学院, 福建 福州 350108
  • 收稿日期:2019-12-16 修回日期:2020-06-07 发布日期:2020-12-25
  • 通讯作者: 涂伟 E-mail:tuwei@szu.edu.cn
  • 作者简介:叶浩宇(1996-),男,硕士生,研究方向为城市时空大数据分析。E-mail:yehaoyuchn@whu.edu.cn
  • 基金资助:
    国家重点研发计划(2019YFB2103104);广东省自然科学基金(2019A1515011049);深圳市科技创新委员会基础研究(JCYJ20170412105839839)

Deep reinforcement learning based electric taxi service optimization

YE Haoyu1, TU Wei2,3,4,5, YE Hehui6, MAI Ke2,3,4, ZHAO Tianhong2,4, LI Qingquan1,2,3,4   

  1. 1. State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, Wuhan 430079, China;
    2. Department of Urban Informatics, School of Architecture and Urban Planning, Shenzhen University, Shenzhen 518060, China;
    3. Guangdong Laboratory of Artificial Intelligence and Digital Economy(SZ), Shenzhen University, Shenzhen 518060, China;
    4. Guangdong Key Laboratory of Urban Informatics, Shenzhen University, Shenzhen 518060, China;
    5. Key Laboratory for Geo-Environmental Monitoring of Great Bay Area, MNR, Shenzhen 518060, China;
    6. Software College, Minjiang University, Fuzhou 350108, China
  • Received:2019-12-16 Revised:2020-06-07 Published:2020-12-25
  • Supported by:
    The National Key Research and Development Program of China (No. 2019YFB2103104);The Natural Science Foundation of Guangdong Province(No. 2019A1515011049);The Basic Research Projects of Shenzhen Technology Innovation Commission (No. JCYJ20170412105839839)

摘要: 作为公共交通的重要组成部分,电动出租车对电动车推广具有重要的示范意义。相较于燃油出租车,电动出租车需要耗费更多充电时间,降低了出租车司机的使用意愿,全面推广面临较大阻力。强化学习方法方兴未艾,适用于出租车运营的顺序决策过程。基于强化学习,本文构建双深度Q学习网络(double deep Q-learning network,DDQN)模型模拟电动出租车的运行。根据出租车的实时状态选择并执行最优载客、充电、空驶和等待等动作,通过训练得到全局最优的电动出租车运营策略,实现电动出租车运营智能优化。利用美国纽约市曼哈顿岛的出租车出行数据进行试验。结果表明:相较于简单的电动出租车运营模式,DDQN优化策略最高将充电等待时长降低70%,拒载率降低53%,司机的载客收入提高7%。相对于电池容量,充电速率和车辆总数对出租车运营收入影响更大,当充电速率达到120 kW时,电动出租车达到最佳的运营表现,政府在推广电动出租车的过程中应当建设更多高速率的充电站以提升电动出租车的运营表现。

关键词: 深度强化学习, 电动出租车, DDQN, 出租车运营

Abstract: Electric taxis have been demonstrated with the promotion of electric vehicles. Compared with internal combustion engine vehicles, electric taxis spend more time in recharging, which reduces the taxi drivers’ intention to use. Reinforcement learning is applicable to the sequential decision-making process of taxis driver. This paper presents the double deep Q-learning network (DDQN) model to simulate the operation of electric taxis. According to the real-time state of taxis, DDQN will choose the optimal actions to execute. After training, we obtain a global optimal electric taxi service strategy, and finally optimize the taxi service. Using real-world taxi travel data, an experiment is conducted in Manhattan Island in New York City, USA. Results show that, comparing with the baseline methods, DDQN reduces the waiting time for charging and the rejection rate by 70% and 53%, respectively. Taxi drives’ income are finally increased by about 7%. Moreover, the results of model parameter sensitivity analysis indicate that the charge speed and the number of vehicles have greater impact on drives’ income than the battery capacity. When the charging rate reaches 120 kW, electric taxis achieve the best performance. The government should build more fast charging station to improve the revenue of electric taxis.

Key words: deep reinforcement learning, electric taxi, DDQN, taxi service strategies

中图分类号: