面向城市人群时空热点预测的混合神经网络  

A Deep Hybrid Neural Network Model Oriented to Urban Crowd Spatio-temporal Hotspot Prediction

作　　者：游兰 张涵钰 韩凡宇 金红 崔海波 何渡[2] 汪坤钰 郑巧仙 YOU Lan;ZHANG Han-yu;HAN Fan-yu;JIN Hong;CUI Hai-bo;HE Du;WANG Kun-yu;ZHENG Qiao-xian(School of Computer Science and Information Engineering,Hubei University,Wuhan 430062,China;Hubei Academy of Scientific and Technical Information,Wuhan 430071,China)

机构地区：[1]湖北大学计算机与信息工程学院,湖北武汉430062 [2]湖北省科技信息研究院,湖北武汉430071

出　　处：《计算机技术与发展》2023年第6期194-201,共8页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61803149);湖北省重点专项(2022BAA044);湖北省教育厅科学技术研究计划重点项目(D20201006)。

摘　　要：城市人群时空热点预测对公共安全应急决策有重要的意义。城市人群热点区域往往伴随时间空间的推移而快速演化,如何发掘利用热区的时空相关性是精准预测城市人群热点变化趋势的关键。该文提出了一种基于深度学习的混合神经网络模型,即CNN-Seq2Seq-Attention(CSA),用来预测连续一周城市人群热点的时空变化分布。为了较好捕捉热点区域的空间信息,CSA模型采用卷积神经网络提取城市热点区域的特征向量,同时,考虑到长时时序数据的周期性,CSA结合Seq2Seq与Attention注意力机制建模人群热点在连续特征日下相同时间片段的时间周期规律。其中,针对人群热点随时间变化的不均匀特性,CSA采用了一种改进的时间片段划分方法,即,基于生活作息的不等长时间段作为数据划分依据。实验使用了连续3个月的城市出租车轨迹数据集,将每周7天标识成7个特征日,每个特征日被划分为7个时间片段,采用预测结果的均方根误差(RMSE)为评估指标。实验结果表明,较传统的PreHA、HA和ARIMA方法,CSA模型效果更好,同时,相较Seq2Seq和CNN-Seq2Seq模型,CSA模型预测误差最大分别降低6.4%和3.8%。The prediction of the urban crow hotspots is of great significance to the public security emergency decision.The urban crow hotspots always evolve rapidly with the changes of time and space.The key to accurately predict the trends of the urban crowd hotspots is how to explore and utilize the spatio-temporal correlations of the hotspots.We propose a hybrid neural network model based on deep learning,namely CNN-seq2seq-attention(CSA)model,for the urban crowd hotspots predictions.Considering the spatial correlations among hotspots areas,the eigenvectors of urban hotspots are extracted through the CNNs model.Also,CSA is combined the Seq2Seq and Attention mechanism to model the time cycle rules of crowd hotspots for the certain time segment in continuous days.Meanwhile,in view of the uneven characteristics of urban crowd hotspots changing with time,an improved time divisions method is designed in CSA,which is an unequal time periods division method based on urban daily schedules.In this paper,the urban taxi track dataset of 3 consecutive months is used in the experiment.7 days per week are identified as seven featured-days,each of which is divided into 7 time segments.The mean square error(RMSE)is the evaluation index.The experimental results show that the traditional methods including PreHA,HA and ARIMA are not as good as neural networks,and CSA can get better accuracy.Moreover,compared with Seq2Seq and CNN-Seq2Seq,CSA can reduce the prediction errors by 6.4%and 3.8%respectively.

关 键 词：社会计算 时空数据 混合神经网络 城市人群热点 时空相关性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]