基于粒子群和LSTM模型的变区间短时停车需求预测方法  

A Prediction Method for Short-term Parking Demands in Variable Interval Based on Particle Swarm Optimization and LSTM Model

作　　者：刘东辉[1] 肖雪 张珏[3] LIU Donghui;XIAO Xue;ZHANG Jue(Jilin Police College Department of Information Engineering,Changchun 130000,China;School of Communication,Jilin University,Changchun 130021,China;Traffic Police Detachment of Hangzhou Public Security Bureau,Hangzhou 310000,China)

机构地区：[1]吉林警察学院信息工程系,长春130000 [2]吉林大交通学院,长春130021 [3]杭州市公安局交通警察支队,杭州310000

出　　处：《交通信息与安全》2021年第4期77-83,共7页Journal of Transport Information and Safety

基　　金：国家自然科学基金青年科学基金项目(51308249);吉林省智能交通创新团队项目(20190101023JH)资助。

摘　　要：停车信息是智能停车诱导系统得以成功实施的关键与基础,被广泛认为能够有效解决当前停车难问题。鉴于停车信息在解决停车问题中的重要性,研究了基于粒子群和LSTM模型的变区间短时停车需求预测方法。为充分发挥数据在提高模型预测精度的作用,提出了以马尔可夫生灭过程为基础概率转移模型,将停车到达率、离开率量化车随时间变化的停车需求,通过标定实际的停车到达率和离开率,确定预测模型的动态预测间隔与时段;采用LSTM网络作为基础预测模型,并利用粒子群优化算法优化网络参数。以吉林大学南岭校区停车场为研究对象,按工作日与非工作日分别对停车数据进行预测并与其他预测模型进行对比分析。结果表明:提出的停车需求预测模型在工作日的预测平均绝对误差为2.53辆,均方误差为11.89辆;非工作日的预测平均绝对误差为2.32辆,均方误差为10.89辆。An intelligent parking guidance system is widely considered to solve the problem of difficult parking at present,providing management,traffic participants,and parking operators with immediate and future parking information.A prediction method for short-term parking demands in variable interval based on particle swarm optimization and LSTM model is studied due to the importance of parking information.Based on the birth and death of the Markov process,the characteristics of the temporal parking demand are analyzed.It is formulated as a combination of the arrival rate and departure rate of parking calibrated by the temporal parking quantity.Dynamic prediction intervals are determined according to the calibrated arrival rate and departure rate.The improved LSTM network is used as the basic prediction model,and the network parameters are optimized by the particle swarm optimization algorithm.The parking lot in the Nanling campus of Jilin University is selected as a research object,and its parking data are predicted and compared with other prediction models.The results show that the MAE and MSE of the proposed parking demand prediction model are 2.53 vehicles and 11.89 vehicles in working days,respectively.For non-business days,the MAE is2.32 vehicles and the MSE is 10.89 vehicles.Therefore,a predictable prediction model of parking demands proposed in the work can predict the real-time and future parking demands,providing a reliable reference for management,traffic participants,and parking operators.

关 键 词：智慧停车 实时及未来时刻停车信息 短时停车需求预测 马尔可夫生灭过程 LSTM 

分 类 号：U491.7[交通运输工程—交通运输规划与管理]