测绘学报 ›› 2018, Vol. 47 ›› Issue (12): 1660-1669.doi: 10.11947/j.AGCS.2018.20170268

• 地图学与地理信息 • 上一篇    下一篇

一种基于模糊长短期神经网络的移动对象轨迹预测算法

李明晓1,2, 张恒才1, 仇培元1, 程诗奋1,2, 陈洁1, 陆锋1,2,3   

  1. 1. 中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室, 北京 100101;
    2. 中国科学院大学, 北京 100049;
    3. 江苏省地理信息资源开发与利用协同创新中心, 江苏 南京 210023
  • 收稿日期:2017-05-22 修回日期:2018-01-03 出版日期:2018-12-20 发布日期:2018-12-24
  • 通讯作者: 张恒才 E-mail:zhanghc@lreis.ac.cn
  • 作者简介:李明晓(1991-),男,博士生,研究方向为时空数据挖掘。E-mail:limx@lreis.ac.cn
  • 基金资助:
    国家自然科学基金(41771436;41571431;41771476);国家重点研发计划(2016YFB0502104);中国科学院重点项目(ZDRW-ZS-2016-6-3)

Predicting Future Locations with Deep Fuzzy-LSTM Network

LI Mingxiao1,2, ZHANG Hengcai1, QIU Peiyuan1, CHENG Shifen1,2, CHEN Jie1, LU Feng1,2,3   

  1. 1. State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China;
    2. University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China
  • Received:2017-05-22 Revised:2018-01-03 Online:2018-12-20 Published:2018-12-24
  • Supported by:
    The National Natural Science Foundation of China (Nos. 41771436;41571431;41771476);The National Key Research and Development Program (No. 2016YFB0502104);The Key Research Program of the Chinese Academy of Sciences (No. ZDRW-ZS-2016-6-3)

摘要: 预测移动对象未来某时刻位置能够为城市规划与管理、城市公共安全、城市应急指挥等提供重要的决策依据,也可为个性化信息推荐、广告定投等基于位置的服务应用提供技术支持。已有预测算法多采用固定格网剖分,位置相近轨迹点常被划分至不同格网,使得潜在轨迹模式被忽略,降低了预测精度。此外,已有预测模型不能有效学习到长序列轨迹有效信息,造成长期依赖问题。本文提出一种基于模糊长短时记忆神经网络(fuzzy long short term memory network,Fuzzy-LSTM)模型的移动对象轨迹预测算法,引入模糊轨迹概念解决固定格网剖分所导致的尖锐边界问题,并对传统LSTM进行改进,综合利用移动对象历史轨迹邻近性和周期性出行特征,提高移动对象轨迹位置预测精度。最后,采用某市10万用户连续15个工作日的移动通讯信令轨迹数据集对方法进行试验分析。结果表明,本文方法在30 min预测周期内的预测平均准确率达到83.98%,较经典的Naïve-LSTM预测模型和NLPMM预测模型分别提高了4.36%和6.95%。

关键词: 位置预测, 模糊空间划分, LSTM, 轨迹数据挖掘, 深度学习

Abstract: Current studies on trajectory prediction have two limitations. Spatial division approaches in most existing studies lead to sharp boundary problem of predicting methods. On the other hand, most of traditional predicting models such as Markov could only use a few latest historical locations, making long-term prediction inaccurate. To overcome these limitations,a location prediction method based on deepFuzzy-LSTM Network is proposed. The method employs a long short term memory based network to solve the long-term dependencies problem. By defining the fuzzy-based trajectory and the improved LSTM cell structure, our method solves the sharp boundary problem caused by space partition. It also considers both period and closeness of movement patterns in making prediction. We compare classical NLPMM, Naïve-LSTM and Fuzzy-LSTM methods with a cell signaling dataset consisting of the continuous trajectories of one hundred thousand city residents in 15 workdays. Results show that the proposed Fuzzy-LSTM method gets a precision of 83.98%, 6.95% higher than the NLPMM model and 4.36% higher than Naïve-LSTM model.

Key words: location prediction, fuzzy space partition, long short term memory network (LSTM), data mining, deep learning

中图分类号: